At an urgent UK parliamentary hearing, OpenAI, Anthropic, Google DeepMind, and Meta split on whether the AI Security Institute should get real powers, with Anthropic calling for binding rules and OpenAI backing mandatory incident reporting. It was a quiet week for verifiable news elsewhere, so this edition stays short.
Under oath-style questioning at the New York City Council, OpenAI, Anthropic, Meta, and Google all declined to guarantee their AI agents will follow guardrails. Anthropic also formalized tiered cyber access for vetted defenders, and OpenAI detailed its Australian government-site incident.
The White House and six AI developers signed a voluntary safety accord with no enforcement, while Anthropic warned that an open-weight Chinese model is close to Mythos-class exploit ability. OpenAI also scrapped a model over safety concerns and fired three safety researchers in the same week.
Twenty nations, OpenAI, and the UN Security Council all made competing pitches for how to govern frontier AI during UN General Assembly week, while security researchers used Claude Opus 5 to hack into OpenAI’s own internal systems. Also: a zero-click bug still exposes two of the four most popular AI coding agents, and Meta and OpenAI both shipped new agent products with very different security architectures.
Anthropic CEO Dario Amodei’s public call to deliberately slow frontier AI development touched off a rare industry-wide argument this week, while California’s governor moved to build an actual ‘kill switch’ for frontier models. Also: a critical Docker flaw let AI coding agents escape their sandboxes on macOS, and the US government’s own benchmark numbers show how close China’s best open model has gotten to the frontier.
Anthropic’s fourth threat intelligence report disclosed that newer Claude models can no longer be assumed safe from meaningfully assisting bioweapons development, and detailed a Russia-linked group that used Claude Code to help build an autonomous, target-selecting drone swarm. Days later, OpenAI reversed course and asked Congress for mandatory AI safety rules, while the FBI, NSA, and CISA warned that Chinese AI firms are systematically distilling US models to catch up for a fraction of the cost.
OpenAI, Google DeepMind, and Anthropic each shipped a frontier model explicitly built to find and exploit security vulnerabilities this week, and each drew its access fence in a different place. Also: a single flaw in how AI coding agents run background Git commands put seven of them one synced folder away from silent code execution, and a federal judge ruled the Pentagon’s blacklisting of Anthropic was unconstitutional retaliation.
OpenAI rallied 128 companies — including Google, Anthropic, and Microsoft — behind a letter warning that AI-enabled cyberattacks are about to outpace status-quo defenses, the same week Google moved its 90-person AI safety team out of DeepMind and into its lobbying arm. Also: Australian police arrested two men behind a supply-chain attack that hit an AI gateway and over 1,000 organizations, and Anthropic showed automated researchers can already improve a model’s own alignment training.
OpenAI spent the week building containment around its own frontier models — a two-week training pause, network isolation, 30-minute alert response, and a zero-data-retention option for enterprise customers. Meanwhile, five US agencies confirmed AI-generated exploits are actively probing Siemens industrial controllers at water and energy utilities, and a Grok vulnerability that lets a web page silently exfiltrate a user’s chat history has sat unpatched since June.
Anthropic raised its own catastrophic-misalignment risk rating and disclosed a biosafety lapse, while OpenAI paused its next model over cyber capabilities it ‘cannot rule out’ are critical — two labs’ own frameworks triggering real brakes in the same week. Around that: joint US-UK testing found a Chinese open-weight model can still complete a full cyberattack chain, Black Hat researchers described criminals turning to open-weight models to lower their barriers to entry, and OpenAI armed vetted defenders with a purpose-built exploit-development model.